EP 89: AI's Role in Responsible Research

This is the Everyday AI show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business and everyday life.

How can we use AI for research without it either taking forever or without it just lying right? You know, there's so many breakthroughs happening that require research papers, but sometimes they can take months or longer. So that's one of the things that we're going to be talking about today on Everyday AI. I'm very excited to join you. My name is Jordan Wilson. I'm your host and we do this every single day. Every day. AI is for everyday people and we're all trying to learn and leverage AI. So that's what we're talking about today.

Excited?

If you're joining us live, make sure to leave us a comment when we bring our guest on. And what, what do you want to know? What do you want to know about using AI for responsible research? Right? Is this something you've seen? Is it something you've experienced? Let us know. If you're joining on the podcast. Make sure to check out the show notes. We will have links to even join the conversation and you can ask our guest questions as well as other just very important tools, tips, things that we're going to be talking about. It's all in the show notes. So let's get started first with AI news. So a major AI breakthrough in heart health. So researchers at Osaka Metropolitan University just released some groundbreaking and accurate AI based methods for classifying cardiac function. So they're using some new AI technology and chest X rays to better and more accurately diagnose heart conditions. So exciting, exciting news there. So make sure to check that out in the newsletter. Not so exciting news. Apparently the AI boom is causing an epidemic of underpaid overseas workers. So a new Washington Post expo to say kind of lift the veils up on this one. So there is a company called Remote Tasks which uses remote workers in the Philippines for different tasks. So this Washington Post report kind of showed that a lot of the workers interviewed for this story were making less than the minimum wage in the Philippines, which is $6 to $10 a day. So scale AI is kind of the owner reportedly of this remote task. So something to keep an eye on. But this is, it's not new. But you know, so many of these different models in AI need a lot of human training. And just kind of this, this piece kind of goes into the uglier side of that. All right, last but not least, journalists, you can keep your job for Now, Right. So a lot of companies have been experimenting with having AI kind of be their writers. Sometimes it works, a lot of times it doesn't. So there was a recent mishap in the Columbus Dispatch using their AI sports writing tool. So this was a recap of a football game at Westerville and it faced a lot of criticism on social media for some of the action packed phrases like close encounter of the athletic kind. Yes, that is what their AI writing tool described this football game as, a close encounter of the athletic kind. Don't worry though, there's people out there working on more responsible AI tools that don't hallucinate or make things up like close encounters of the athletic kind. So let's actually bring our guest on for today and talk about how we can use more responsible AI tools to get better results. So welcome to the show. Avi Saman, the founder of Cyrider AI. Avi, thank you for joining us.

Thanks so much, Jordan. It's a pleasure being here. And from that headline, it sounded like maybe a UFO kind of landed in the middle of the football game. So, you know, who knows, maybe that was an accurate description.

But you're right. And I, I guess that could be athletic, you know, if people had to run and jump out of the way. That's good. So, hey, as a reminder, if you're joining us live, like Mercy is saying, hi, everyone, make sure get your question in for Avi. What do you want to know about responsible AI research? Val saying good morning, woozy. Good morning, Dr. Muthana. We'll get to your question there in a second, but thank you everyone else for joining us live this morning. Avi, what's wrong? What's going on right now with AI and research tools? And what led you to. Because I'm sure you saw things that were going wrong or things that just weren't right out there in the field. So kind of like what led you to create Sci Writer and why is it needed?

Yeah, so that's a great question, Jordan. First of all, I would say that I think all of us, anyone who's played around with ChatGPT or any of the generative AI tools has come across a hallucination. And hallucination, otherwise known as crap that's made up, that isn't accurate. And I think sometimes we forget that the second L of LLM stands for language. And we, we kind of treat it as a, I don't know, Wikipedia lite sort of tool and try to get all information from there, where actually it's purporting to be a language tool. So the issue in my specific field, which is scientific research and the publication of that research, is that we can't afford a high degree or even a low degree of hallucinations if we're just writing marketing copy. So maybe we can get a good first draft and we can play around with it, fix up the mistakes and send it off. If we're talking about doctors who in real time are looking to the scientific record, looking to research to answer critical questions on the fly, well, then they can't afford to have those mistakes that are baked in and built in. So I think there's an, in general, there's a trend, there's this sort of small section which maybe isn't known as much, but research tools that are using AI to kind of harness the good of generative AI and use that as a tool. So cy Writer specifically, which is the project that I'm working on with a buddy of mine, is an attempt to ask ourselves, can we take the power of, let's say, a writing tutor? Right. Anyone who's ever worked with a writing tutor before knows it's can be so powerful to have these. This question and answer this dialogue, someone to help you actually tease out what you want to say. Well, can we turn, can we use generative AI and can we ask the researcher, Give us, tell us what methods you used. Give us your, you know, the results from your lab or from the library that you were doing your research in. Feed that to us and then we can take that and turn it into an output that actually resembles what a typical article is like and bring the hallucinations down to nearly zero. And if we can do that, then what we're actually doing is saving researchers a lot of time. And that translates into more time for them to be doing their cancer research, for them to be tutoring their students so that they're the next generation of researchers, for them to be explaining their research to the public. So that's why we see our tool as really critical for the next generation of research and science.

Yeah. So talk avi a little bit about what this process looks like now for researchers, because it sounds like it's something you eat. Either run. Yeah. If you want your research out quickly, you either run the risk of, you know, maybe using some other, you know, like a chat GPT type tool out there and getting hallucinations, or it just might take forever to, to get a new, you know, scientific breakthrough out to the masses. So, so what is it like now for researchers to, you Know, is there just too much gray area on, hey, what's the right way to maybe tap into AI to expedite that process?

Yeah, the answer is it's a big mess. First, I mean, I can. I think we can all relate back to a couple years ago when we were in the, you know, climax, the pandemic, and we were all waiting for these labs to come out and say, okay, here's, you know, kind of the latest study, here's the vaccine, here's what we're proposing. And I think everyone was kind of frustrated by the pace, even when it was expedited and even when the drugs were pushed through. To have that peer review, it's really critical to go through what's called the peer review process. The peer review process is essentially what takes a study that someone purports or a claim that someone purports to have made based on their research and turns it into actually verifiable scientific literature that we all rely on for all sorts of decisions on a daily basis. So now the issue becomes that process can take two to three years. And I think in the pandemic, we realized, like, holy crap, we don't have two to three years to actually do this. So what generally happens, I can tell you, like a typical scenario, is that the researchers do the study. The study itself could take anywhere from six months to two or three years, depending on what you're actually studying. Then comes the writing stage. And during the writing period. So oftentimes the researchers who are running these labs are running a bunch of experiments and trials in parallel, and they don't really have the time to write up the research. So they'll give it off to a student and they'll say, you know, here, master's student, PhD student, you know, go and write this up for me. We've already done the research. How hard could it be? It's really not easy. They break their teeth, they struggle. They get super frustrated. In fact, half of the authors that publish, half of the students and researchers that publish never publish again because it's such a frustrating experience. It's a really big problem. Then they send it back to their professor who's running the lab. Professor's like, oh, my goodness, this is total rubbish. This is crap. I need to throw it out and start over. And then they rewrite it, and only then it gets sent to the publisher, and then I won't bore everyone with the gory details. But even at the publisher, there can be endless back and forth with the publisher about what's accepted? Does it meet their formatting standards? Does it meet their, their, you know, can it be reproduced? So this whole process is very clunky, and I think part of it is important. We don't want someone to submit an article and the next day it's published because then we're like, whoa, did anyone actually look at this? Like, are we okay with that? That's a problem. If it takes two years, it's also a problem. So my goal is to kind of boil that time down and experience, first of all is to bring the time down considerably, but second of all is to turn that experience into enjoyable. This should be the climax. You finished your research, you want to tell the world about what you've discovered, and then you just have this downer experience. So that's what we're trying to do at Sci Rider is really turn that around. Yeah.

And we actually have a couple people in the medical field here joining us in the comments. So, Dr. Harvey Castro, thanks for joining us. A good, a good question here, Avi. So this, this one might go over my head, but Dr. Rastafa get us asking, you know, about oral defenses. So I'm guessing, you know, that's part of the process of the research, you know, getting something out. And then so he's kind of asking like, is there a good way or a bad way to even maybe use different AI tools to help in that oral defense, I'm guessing is part of the process to get this out there. I'm not sure. But, you know, walk us through and maybe whether it's Sci Writer or if there's other AI tools that maybe are or aren't a good idea in that step of the process.

Yeah. So I'm not sure that I fully understand exactly what he's referring to, but I can tell you when it comes. I have seen a number of researchers and teachers at universities and colleges who have said that they are going to actually phase out written works because they're afraid of AI generated, you know, works, and they're going to rely quite heavily on oral presentations and defenses. I'm not, I don't think we should be running to do that because I think there's a power to the written word that it can be shared afterwards and then critiqued and followed up on and revised, whereas an oral presentation doesn't exactly give that. I also think that why are we trying to look away from generative AI? Can we ask the students to think critically about generative AI and, and actually teach them how to prompt in a way which can get them the best results. So I'm not, I'm not against, you know, kind of, I think it's actually really critical that researchers learn how to speak their research and not just how to write their research or how to understand it themselves. And a lot of researchers are terrible at that. But I'm not sure that I would use that as a replacement for the written record. I think it's important that anyone, anywhere at any time can actually tap in and be able to look up, you know, if God forbid, I don't know, your grandparent has some illness. And you want to understand, you're not a doctor, you're not a researcher, you understand what's going on here. There are amazing tools that are being built now to create layman summaries, right? So that me and you as non subject experts can go in and understand what the heck do they have, what are the treatment possibilities, how, you know, what, what doctors should I be looking to and how do I ask those questions. So that's kind of where I'm excited about some of the specific research tools in the AI space.

You know, a couple, couple of questions and comments here. I want to get to one more. Avi. So you know Bronwyn, just saying this would have been helpful when I had to research stem cell therapy kind what you said, even for people maybe who aren't, you know, trying to publish. This is something that can help just people understand topics.

For sure.

Brian, Just saying accuracy is paramount in social science research too. Hallucinations definitely make you look incompetent. But here's, here's a question Avi, I'd like your take on Dr. Muthana is asking. So you know, we're just talking about that a lot of different publishers out there are now blocking large language models from seeing the information out on their website. So you know, let's say there's, there's large, you know, companies that normally put out great scientific research. So what happens for the large language models out there? And maybe, you know, not specifically sci writer, but maybe, maybe so, so what happens when all these publishers start blocking access, you know, to these large language models to get this, you know, really needed information that would in theory help make that model smarter on whatever scientific research someone is using it for.

Yeah, this is a really good question. So I want to break it down to a few parts. First of all is it's not clear what's been used already and what hasn't been used, what's been scraped and what hasn't been scraped right and what's interesting about the academic publishing industry is that they have a lot of things behind a paywall, right. And they've been doing this for many years. So imagine, I don't know, you want access to New York Times article and like all of a sudden you get stuck. Well, actually that content is super valuable and it may not be so simple for OpenAI to use that even things that are available might not be under a license that can be used by these AI companies. So it's unclear exactly what they have, what they don't have and what they're missing. My claim, and I recently wrote about this in a magazine called the Scholarly Kitchen, which I'm also a member of the editorial board. There is that actually there's a tremendous opportunity here because if we think about after all the hype dies down and after we all kind of buckle up and say, okay, what can we actually do with generative AI? I think that for the majority of use cases, we're going to want it to be relying on reputable, verified, important information. Okay. And with all due credit to Reddit, I don't think the regurgitating Reddit does that much good for society. It may help understanding kind of how people talk and maybe replicate that, but it's not going, no one's going to take that and put it into their doctor's office. So I think the real question here is how can academic publishers who own and basically can license this content and the large language models, how can they join forces to actually cover an entire field, cover an entire space, not just what's available through Wikipedia or you know, secondary sources, but actually go back to the original research and then build either large language models or what I've seen is more small language models that are very hyper specific based on certain content inputs and actually take that and turn it into life saving applications. That's where I get really exciting about it, excited about it. It's a challenge because first of all, different publishers have different content. So you mentioned at the beginning of the show about cardiology. Well, you may have some of that content might be with one publisher and some of it might be another. So you need to get these publishers to work together, then you need them to trust the AI company they're working with. So I think it's beginning and starting and I think there's probably a lot going on behind the scenes, but it's still early days so we'll have see how this develops.

Yeah, and speaking of developing, I think there's a lot of just developments in general. Right. Because now you also have bigger companies, you know, like Google, you know, trying to be a player in this space. So, you know, with their new. I think Palm 2 is kind of a large language model for the, for the medical field. So in general, and maybe not just specifically talking about, you know, Palm2avi, but is it a good thing or a bad thing when you have, you know, large companies like Google, you know, creating models specifically for, you know, the medical space? Right. So that's obviously what a lot of scientific, you know, research articles, that's, that's kind of their category, so to speak. So is that a good or a bad thing? And does that make it, you know, easier or maybe more difficult for, for researchers to have something like Google Palm as a resource?

Yeah, so I think that they. To use it as a resource is perfectly fine and I think even recommended. I want to warn of two issues that I've already seen that can be really damaging if you don't know how to use it properly. Example number one, and these are both on the one hand funny, but also on the other hand, sad and ironic. The first example was a professor. He got a bunch of student, student works or student assignments that he got back papers at the end of last year's semester. And he then was, you know, wanted to be very vigilant and decided that he wanted to check if they were written by generative AI or not. And he threw it. He actually put the papers into ChatGPT and asked GPT, well, did you write these papers? And GPT says, yes, of course I did. And not realizing that that's not a great question to ask ChatGPT. And he then went on to fail all of his students. And the students basically had, you know, had a mini rebellion because they're like, we didn't use ChatGPT. So that's kind of one example which again is comical, but also quite sad. The second example, which is even more troubling potentially is if you go into Google Scholar and Google Scholar is kind of like, you know, to find research articles through Google and you, I believe it's. You type in the name of like, you know, academic articles and then, and then specifically look for the words regenerate response. You will find there is already published articles that have the words regenerate response in there. Now I will give them the benefit of the doubt. There may be that phrase may exist in certain academic articles, but most likely that's a sign that these researchers, not only do they copy and paste directly from ChatGPT they may not have even gone over the output and may have just published as is. Even worse, a lot of the stuff is actually published. It's not just that someone threw it up on a blog, anyone can do that, but it's actually made it through that peer reviewed process which is supposed to catch those issues and errors before they happen. So I just think it's, you know, I'm always like in two minds because the kind of entrepreneur, innovator, you know, embrace the good of technology and me says, yeah, like let's use it. Like if you're in, for example, if you're not a native English speaker, right, and you need to publish your article and you're competing against your American colleagues and your British colleagues and your Australian colleagues, and you find writing English to be really hard, it could be your third or fourth language and maybe you never learned it in school. Using GPT to edit your work is, is, is really, really can be game changing. It can be totally game changing or to help you to organize your references according to a specific style guide. So on that part of me says, yeah, like let's embrace it. And then the other part of me when I hear these stories is like, oh, that's cringeworthy, right? Like, like this is really problematic. And if these are the stories we know about, what about the stories we don't even know about? So I think the answer is education. Education. Doing podcasts like these, getting out and explaining how to use it right, how to not use it right, and making sure you're part of that conversation and dialogue. You're not afraid of it on the one hand, but you're also not just using it blindly on the other.

Are you still running in circles trying to figure out how to actually grow.

Your business with AI?

Maybe your company has been tinkering with large language models for a year or more, but can't really get traction to find ROI on Genai. Hey, this is Jordan Wilson, host of this very podcast. Companies like Adobe, Microsoft and Nvidia have partnered with us because they trust our expertise in educating the masses around generative AI to get ahead of. And some of the most innovative companies in the country hire us to help with their AI strategy and to train hundreds of their employees on how to use Genai. So whether you're looking for ChatGPT training for thousands or just need help building your front end AI strategy, you can partner with us too.

Just like some of the biggest companies in the world do.

Go to your everydayai.com partner to get in contact with our Team or you can just click on the partner section of our website will help you stop running in those AI circles and help get your team ahead and build a straight path to ROI on gen AI.

Yeah, and I think, Avi, that's a good point because so many people will just blindly, you know, use, you know, something like ChatGPT or Google's Bard and you know, kind of think of it as fact, right? And they say, all right, well hey, if you know, Google Bard or Microsoft Bing Chat gives me this response. It must be, you know, ready to go. It must be well researched. It must be, you know, ready for whatever I'm going to be using it for. So kind of like with that in mind, I know we've talked about a lot top to bottom on the show so far, but what's kind of your, your one piece of advice for people out there, whether they're reading, you know, recent, you know, scientific studies or maybe they're, they're writing them, what is the one piece of advice that you have for people to responsibly use AI, kind of regardless of where their, you know, where their output may be.

So the way I describe it is in three words, Wordle on steroids. Okay. That's what I call ChatGPT. For those who aren't familiar, Wordle was this game that got really popular about a year ago where you get, you know, you have to guess five letters and, and make up the word. And why do I think that ChatGPT is Wordle on steroids? First of all, what I said before, it is not Wikipedia light. So let's get that into our minds. Now. There are certain interesting research applications where you could actually add on references. So Scholar AI is a really cool app where you can kind of add. It's an add on to ChatGPT where it gives you actually referenced answers. Really cool. But just GPT on its own. And some of the other generative models are not referenced as we know references are often made up, which is a big problem. So we need to not think of it as an information machine. That doesn't mean that it never has accurate information. It can. But if our expectation when we go into Wikipedia is that we're getting accurate facts, that should not be our baseline assumption. When we're going using a generative AI model, we should be using it for exactly what it says. Language, large language model. So language can be used in all sorts of ways. It can be used to take a very long text and shrink it into a shorter one. It can be used to Take a text and translate it to a different language, which is an application that I'm working on now in my, in my original business, academic language experts. It can be used to even kind of generate, you know, a social post about a certain piece of research that was published. So. But those are all kind of, it takes ideas, it takes words and it processes them, or it can even generate words in a very creative way. But when we think about it as a word language tool and not as an informal source of information or a creator of information, then I think we can be much better empowered. Right? And that professor that I mentioned before would never have asked GPT if he had written this paper or not, because he's not a source of information. Had he asked GPT, is this written? Well, according to the average, you know, standards for a student at, you know, a university after a certain amount. Well, actually GPT might have been able to give him a, you know, a semi intelligible answer. I think part of the blame, by the way, rests with OpenAI. I think they kind of came out and are like, here, take this and figure out what it does. And I think it's important. And I haven't seen enough educational materials on there. And I know it's not just OpenAI, right? There's other Google and Facebook and Anthropic and Hugging Face, but from all these companies, I haven't seen enough educational resources and material around. Well, here is what it can do, here's what it can't do, here's responsible use and not irresponsible use. So I'd love to see more of that. So but in the meantime, let's educate ourselves so that we're using in the best way possible.

Yeah, and I think that's a good point because Avi, you kind of mentioned there's almost too many new tools, almost too much kind of new software out there claiming to help. It seems like there's either a new large language model popping up weekly or at least a large update. So actually, real quick, just, just give a plug. We're going to share it in the newsletter today, so make sure to go to your everydayai.com Sign up for the newsletter because Avi's been dropping a lot of good names. You know, he said like Scholar Scholarly AI or Scholar AI, which is a great ChatGPT plugin for research. So make sure to go sign up for the newsletter and we'll share it with you. But just real quick, tell people about kind of Tool up Tuesday and a great resource that you've Created a great free resource which we're big on here at Everyday AI, but a great resource for people to kind of help write better research papers and know what tools are out there to help them.

Yeah. So I put together AI Tool Up Tuesdays on a whim. I was inspired by a colleague from eo, from Entrepreneurs Organization who had done this in the marketing space. And basically the idea is, as I said, okay, there is a. If we try to look at all the AI tools in the world, well we get overwhelmed very quickly. But in my specific industry, so that's academic research, we actually can, there's, you know, a couple dozen AI tools which I think are mature enough that can actually be used. Some of them are in the research discovery world, some of them are in the research processing or image, you know, image production, like creating new images from scratch. But specifically for science in the writing. That's where Scott, where that's where Sci Writer lies is in the writing tool section. So there's all sorts of different areas where AI can actually be really, really important and helpful. So what I did was I put together AI Tool Up Tuesdays, which is, which has, I guess I'd call said gone viral. We already have over 3,500 researchers from around the world who are actually registered for the course. It's eight sessions. Each session is three entrepreneurs. Most of them are academics, former academics who have built tools to really address real big problems in the academic workflow and in research. So like this month you're showing on the screen, we're gonna be talking about research veracity and integrity. Super critical. How do we make sure that research is actually can be relied upon in a generation and in a time where there's so much, you know, kind of rubbish being flooded out there. So like I said, entirely free. Each entrepreneur researcher only has 10 minutes to present. So it's really, really straight to the point, hard hitting exactly what the capabilities of their tool are. And I think that anyone who's involved in research in any way in their business or in their personal life, this is going to be a can't miss because this will save you hundreds of hours. And I'm not exaggerating in kind of the old fashioned way of writing a doctoral thesis over five, six, seven years into being a super frustrating process. We want this course to really supercharge research. And like you said, it's free, so nothing to lose.

Awesome. Cool. And yeah, as a reminder, make sure to check out the newsletter. There was a lot of great information that Avi shared today and like Dr. Mathana here just just left in the comments. The newsletter does go beyond just just recapping the show. We have a lot of great insights and information in that newsletter and we will include what Avi just talked about the tool Tuesday. So Avi, thank you so much for joining us on today's show. We went all over the place top to bottom, but I think it was important to talk about just responsible AI and research. So thanks again for joining us.

Awesome. Thanks Jordan. If anyone wants to, you know, be in touch or follow LinkedIn's A Great Place. Avi Stamon just my name. Feel free to send out an invite. Quite active on there and sharing my latest thoughts about AI in research. And Jordan, thanks to you for this awesome podcast and really just making the best of AI available and educating us about, you know, what's possible.

Absolutely. So what else is possible? Find out. This week we've got a great, great lineup. I think we have five speakers this week every single day, Monday through Friday. So join us again 7:30am Central Standard Time live. Ask questions just like you, just like y' all did today with experts. We bring on the show with Avi. Avi, thank you again so much. And we'll see you back again tomorrow and every day with every day. AI. Thanks y'.

All.

And that's a wrap for today's edition of Everyday. AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going for a little more AI magic. Visit your everydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.